Production incident root cause analysis powered by Cognee Knowledge Graphs, Vector Search, and LLM reasoning.
🔗 Live Demo: Try TraceMind AI Live Here
TraceMind AI helps engineers investigate production incidents by combining semantic memory, vector similarity search, and knowledge graphs to identify the most likely root cause of system failures. Instead of manually correlating logs, deployments, and alerts, TraceMind automatically builds a contextual timeline and explains why an incident occurred.
Before running any query, your raw infrastructure footprint must be loaded into the system. TraceMind AI processes this using a structured database seeding phase:
- Local Text Vector Generation: A standalone pipeline feeds a mock system history containing GitHub Pull Requests, Prometheus metrics, and Slack communications. A local
FastEmbedmodel processes the descriptive text chunks into vector representations. - Supabase Population: These records and their corresponding multi-dimensional vectors are loaded into a PostgreSQL database using
pgvector. - Cognee Cloud Graph Extraction: An extraction runner pipeline pulls records sequentially from Supabase, streaming them to
Cognee Cloud. Backed by aGroq Llama 3model, it creates links across application profiles, infrastructure layers, and service dependencies to yield a connected, searchable Knowledge Graph.
Runs two independent investigation engines in parallel:
- Correlation Engine – Uses vector similarity and temporal analysis to identify related historical incidents.
- Cognee Reasoning Engine – Traverses the knowledge graph to explain causal relationships between deployments, alerts, and services.
Both results are displayed side-by-side, allowing engineers to compare mathematical evidence with semantic reasoning.
- Fast similarity search using pgvector
- Maps historical production incidents
- Identifies related deployments and pull requests
- Uses temporal decay scoring to prioritize likely root causes
Production events are automatically transformed into a connected knowledge graph containing:
- Application logs
- Git commits
- Deployments
- Pull Requests
- Incident timelines
- Team discussions
- Infrastructure relationships
This enables graph-based reasoning instead of traditional keyword searches.
Beyond incident investigation, TraceMind continuously evaluates system health by analyzing:
- Recent code mutations
- Historical failures
- Service dependencies
- Architecture relationships
The platform predicts high-risk services before production failures occur.
For every investigation, TraceMind provides:
- Root cause explanation
- Risk score
- Related historical incidents
- Suggested mitigation steps
- Engineering recommendations
| Layer | Technology |
|---|---|
| Frontend | React + Tailwind CSS |
| Backend | FastAPI |
| Vector Search | PostgreSQL + pgvector |
| Knowledge Graph | Cognee Cloud |
| Embeddings | FastEmbed |
| LLM | Groq (Llama 3.3 70B & Llama 3.1 8B) |
| Deployment | Render |
Before running the project, install:
- Python 3.11+
- Node.js 18+
- PostgreSQL (or Supabase PostgreSQL)
- Git
cd backendpython -m venv venv
source venv/bin/activatepython -m venv venv
venv\Scripts\activatepip install -r requirements.txtCreate a .env file inside the backend folder.
DATABASE_URL=your_postgresql_or_supabase_connection_string
GROQ_API_KEY=your_groq_api_key
COGNEE_SERVICE_URL=your_cognee_cloud_url
COGNEE_API_KEY=your_cognee_cloud_api_keyFirst, run the database seeding file to load historical alerts, Slack logs, and commits into Supabase with vector representations:
python seed_database.py
Run the ingestion script to load the events from Supabase into Cognee.
```bash
python cognee_ingestion.pyThis process:
- Reads production records
- Creates vector embeddings
- Stores semantic memories
- Builds the knowledge graph
uvicorn main:app --reloadBackend will be available at
http://localhost:8000
Swagger documentation:
http://localhost:8000/docs
cd frontendnpm installCreate a .env file.
VITE_API_BASE_URL=http://localhost:8000npm run devFrontend will run at
http://localhost:5173
TraceMind AI is designed as a Render Monorepo with independent frontend and backend deployments.
| Property | Value |
|---|---|
| Service Type | Static Site |
| Root Directory | frontend |
| Build Command | npm ci && npm run build |
| Publish Directory | dist |
| Property | Value |
|---|---|
| Service Type | Web Service |
| Root Directory | backend |
| Build Command | pip install -r requirements.txt |
| Start Command | uvicorn main:app --host 0.0.0.0 --port $PORT |
Clone the repository and launch both services.
git clone <repository-url>
# Backend
cd backend
python -m venv venv
# Linux/macOS
source venv/bin/activate
# Windows
# venv\Scripts\activate
pip install -r requirements.txt
# Configure .env
python cognee_ingestion.py
uvicorn main:app --reload
# Frontend
cd ../frontend
npm install
# Configure .env
npm run dev- Authentication & user management
- Kubernetes deployment
- Multi-tenant incident memory
- Slack and Microsoft Teams integration
- GitHub webhook automation
- Real-time risk monitoring
Built with ❤️ using FastAPI, Cognee Cloud, PostgreSQL, pgvector, Groq LLMs, and React.
In compliance with The Hangover Part AI Hackathon rules, AI coding assistants (claude/gemini) were used during development to generate boilerplate code, format scripts, and accelerate UI development under our architectural guidance.